Physics-first machine learning for gravitational-wave discovery
My research focuses on physics-first machine-learning approaches to gravitational-wave data analysis, with applications to both ground-based detectors (LIGO) and future space-based missions (LISA). I develop template-free and unsupervised methods designed for robustness, interpretability, and discovery.
Signals without templates
I build anomaly-detection pipelines that do not depend on large waveform banks, making them better suited for discovery-oriented searches and unexpected morphologies.
LIGO to LISA
My work spans both current ground-based detectors and future space-based missions, with an emphasis on the different statistical and computational challenges each regime creates.
Geometry and interpretability
I am especially interested in latent-space structure, invariance, and physically meaningful representations that make machine-learning systems easier to trust and analyze.
Latest Projects
Gauge Freedom and Metric Dependence in Neural Representation Spaces
Neural network representations are often analyzed as vectors in a fixed Euclidean space, even though their coordinates are not uniquely defined. This project treats hidden representations geometrically, as vector spaces defined only up to invertible linear transformations. Within that framework, commonly used similarity measures such as cosine similarity become metric-dependent quantities whose values can change under coordinate transformations that leave the model function unchanged. The result is a common interpretation for observations such as cosine-similarity instability, anisotropy in embedding spaces, and the appeal of methods like SVCCA and CKA. Experiments on multilayer perceptrons and convolutional networks show that invertible transformations can substantially distort cosine similarity and nearest-neighbor structure while leaving predictions unchanged, suggesting that representation analysis should prioritize gauge-invariant quantities or explicitly chosen canonical coordinates.
Global Structure in Learned Latent Representations of Confusion-Limited LISA Data
Machine learning methods in gravitational wave data analyses depend on the choice of representation and on how structure within that representation is used. Building on previous work using continuous wavelet transform (CWT) autoencoder representations for confusion-limited LISA simulation, we investigate whether source resolvability information is better characterized by local latent geometry or by global latent density. We study this question in a controlled benchmark with data generation and preprocessing held fixed. Using CWT representations of synthetic confusion-limited LISA segments, we compare geometry based one-class scoring with likelihood-based latent models along with their morphology augmented variants. Likelihood-based scoring consistently outperforms local manifold-distance methods across three independent seeds, achieving ROC-AUC 0.8555 ± 0.0181 and PR-AUC 0.9219± 0.0118, compared with ROC-AUC 0.7663± 0.0450 and PR-AUC 0.8667± 0.0255 for the geometry baseline. These results suggest that resolvability information in learned latent representations is not fully captured by local latent geometry but instead reflects global properties of the latent distribution. More broadly, this work contributes to representation-aware methods for confusion-foreground characterization in LISA and motivates future studies of coordinate invariance and intrinsic geometry in learned latent spaces.
Detectability Scaling Laws for Environmental Phase Modulation in Gravitational-Wave Signals
Environmental effects such as hierarchical triple motion can introduce cumulative phase modulation in gravitational-wave signals through time-dependent line-of-sight acceleration. This project studies detectability in a template-free framework using continuous-wavelet-transform time-frequency representations and trajectory-based statistics, especially the evolution of the power-weighted frequency centroid. Detection performance collapses onto a single scaling parameter given by phase distortion times signal-to-noise ratio, with ROC-AUC following a sigmoid transition. The main takeaway is that smooth environmental phase modulation is not generically absorbed by intrinsic waveform variability; detectability is governed by a simple scaling between cumulative phase distortion and signal strength.
Manifold Learning for Source Separation in Confusion-Limited Gravitational-Wave Data
The Laser Interferometer Space Antenna will operate in a fundamentally different data-analysis regime than ground-based detectors such as LIGO: instead of rare signals buried in instrumental noise, LISA observations are expected to be dominated by a dense superposition of unresolved Galactic binaries. In this project, I investigate whether manifold-learning techniques can aid source separation in that confusion-dominated setting. I develop a convolutional autoencoder trained exclusively on synthetic confusion-background data and augment the standard reconstruction-error anomaly score with a geometric term derived from the local structure of the learned latent-space manifold. Tests on synthetic datasets with injected massive black hole binaries, extreme mass ratio inspirals, and individual Galactic binaries show that incorporating latent-space geometry substantially improves source discrimination compared with reconstruction error alone.
Manuscript submitted to Classical and Quantum Gravity.
Template-Free Gravitational Wave Detection with CWT-LSTM Autoencoders: A Case Study of Run-Dependent Calibration Effects in LIGO Data
Gravitational-wave searches traditionally rely on matched filtering against large banks of theoretical waveforms, which can be computationally expensive and inherently biased toward known signal morphologies. In this project, I develop a template-free, unsupervised detection framework that combines continuous-wavelet-transform representations with sequence-based machine learning. The method trains an LSTM autoencoder exclusively on detector noise so that gravitational-wave signals appear as anomalies without waveform templates or labeled training data. During development, I found that training across multiple LIGO observing runs caused the latent structure to cluster by observing run rather than by astrophysical signal properties, revealing systematic batch effects tied to calibration and preprocessing. A per-run training strategy eliminated those effects and substantially improved detection performance on O4 LIGO data.